feat: production readiness improvements (P0 + P1)

[DavisSylvester]

• Add ILlmFactory interface to abstract LLM providers (Ollama/OpenAI/Anthropic)
• Add LLM_PROVIDER env var for provider selection with per-provider model maps
• Add token/cost tracking per LLM call with per-model and per-task breakdowns
• Add per-task cost ceiling (TASK_COST_LIMIT, default $3) that aborts runaway tasks
• Add API key redaction in Winston logs (strips sk-, sk-ant-, known key fields)
• Add typed error classes (RateLimitError, ContextWindowExceededError, etc.)
• Add exponential backoff with jitter for retryable LLM errors
• Add --dry-run flag to show task plan without executing
• Add --verbose and --quiet flags for log level control
• Add colored CLI output via picocolors (green/yellow/red levels, cyan phases)
• Add Phase 4 run report generation (report.md in workspace)
• Add end-of-run cost summary (tokens, calls, dollar cost)
• Make package npm-publishable (bin field, version, keywords, files)
• Add GitHub Actions CI workflow (type-check on PR)
• Add CONTRIBUTING.md, SECURITY.md
• Add getting-started guide, bookmark API example PRD
• Add competitive comparison and future improvements docs
• Change OLLAMA_HOST default from private IP to localhost
• Accept raw text and stdin input in addition to file paths